Transfer Learning Algorithms for Plant Classification
Journal of Eye-Innovation in Machine Learning · 12.08.2026 · 1 (20261) · 88
DOI: 10.68990/jieml.202615 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Eye-Innovation in Machine Learning, online erschienen am 12.08.2026. Verfasst von Nizar Alsharif und 1 weiteren Autor:innen. Die Arbeit nennt 46 Literaturangaben.
Abstract (Originalsprache)
The dataset remains useful only after conflict adjudication, quality criteria, duplicate-group isolation, and a frozen leakage-safe split. Literature-reported model scores are contextual evidence and are not comparable to this audit. A CPU smoke test verified software execution only and is deliberately excluded from the performance results. Public plant-image datasets have enabled rapid evaluation of convolutional and transformer-based classifiers, though image duplication and contradictory labels can invalidate seemingly strong benchmark results. The Plants Classification release dataset consists of 30,000 images of 30 nominal classes. We evaluate whether the provided split is suitable for direct performance benchmarking and provide a reproducible, leakage-free classification protocol. The scientific contribution is a completed data-integrity experiment, not a claim that confirmatory neural-network training has been completed. SHA-256 pinned version 1 of the archive, decoded image by image, and described by partition, class, format, color mode, and dimensions. Grouped exact duplicates by file SHA-256. Using identical 64-bit perceptual hashes, we conservatively screened for visually identical or very similar content without performing a full Hamming-radius search. We searched for groups crossing partitions and label contradictions. We then designed a group-stratified 70/10/20 benchmark protocol, excluding contradictory exact-hash groups pending botanical adjudication, with training-only augmentation, ImageNet-pretrained baselines, macro and class-balanced evaluation, and bootstrap confidence intervals. All 30.000 files were decoded, but the image quality was heterogeneous, with dimensions ranging from 1 by 1 to 4.928 by 5.000 pixels, with three images being 1 by 1 pixels. The audit identified 2,567 groups of identical duplicates, comprising 5,554 images. Of these, 1,178 groups crossed supplied partitions and 2,689 images. Moreover, 946 groups with exact-hash had contradictory labels, with 2,052 images. 9,087 groups of identical perceptual hashes were formed, with 21,554 images. 2,936 groups spanned partitions, including 8,285 images. Several partitions contain byte-identical content and sometimes incompatible labels. The original split is not valid for direct benchmarking. A frozen leakage-safe split, adjudication, quality criteria, and duplicategroup isolation are necessary before the dataset is useful. The literature scores for the model are contextual evidence and not comparable to this audit. CPU smoke test that only checked software execution is excluded from performance results by design.
Zitieren
APA
Alsharif, N., & Ebrahim, N. (2026). Transfer Learning Algorithms for Plant Classification. Journal of Eye-Innovation in Machine Learning, 1(20261), 88. https://doi.org/10.68990/jieml.202615
Vancouver
Alsharif N, Ebrahim N. Transfer Learning Algorithms for Plant Classification. Journal of Eye-Innovation in Machine Learning. 2026;1(20261):88. doi:10.68990/jieml.202615
BibTeX
@article{Alsharif2026,
author = {Alsharif, Nizar and Ebrahim, Nadhem},
title = {Transfer Learning Algorithms for Plant Classification},
journal = {Journal of Eye-Innovation in Machine Learning},
year = {2026},
volume = {1},
number = {20261},
pages = {88},
doi = {10.68990/jieml.202615}
}
EUVIMED